APCC: Agile and Precise Congestion Control in Datacenters

Abstract

Modern datacenter networks exhibit complicated and time-varying traffic patterns, from long-running flows to burst short-lived flows. Recently in-network-telemetry (INT) is employed by datacenter transports to perform precise congestion control (CC). Current INT-based CC works well for long flows, however, suffers from serious performance downgrades when BDP-level small flows burst in the first RTT, due to the reason that current INT information from the receivers (host-based INT) needs at least one RTT to respond. In this paper, we make the first attempt to propose an agile and precise congestion control, called APCC, in datacenter networks, working for traffic patterns that is a mix of BDP-level short flows and long flows. APCC explores INT information from switches (switch-based INT) to feedback the congestion information eagerly, and effectively manage BDP-level flows. APCC utilizes the switch-based INT to schedule the complicated and time-varying traffic stably and precisely, and achieve low latency, high bandwidth and network stability simultaneously. We conduct extensive experiments to evaluate the performance of APCC. The experiment results show that with data centers load containing many BDP-level flows such as Cache Follower and Web Server, switch-based INT shows huge potential of improving the long tail effect on completion time (FCT). APCC reduces tail delay by 21.7%-28.9% under normal circumstances, and can still reduce tail delay by 9.1% under more severe conditions comparing with HPCC. Moreover, APCC shows better convergence.